Skip to content

MFIBA: Multiscale Feature Importance-based Bit Allocation for End-to-End Feature Coding for Machines

License

Notifications You must be signed in to change notification settings

SYSU-Video/MFIBA

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

24 Commits
 
 
 
 
 
 
 
 

Repository files navigation

MFIBA

Multiscale Feature Importance-based Bit Allocation for End-to-End Feature Coding for Machines

Python PyTorch arXiv License Last commit

Multiscale Feature Importance-based Bit Allocation for End-to-End Feature Coding for Machines
[paper] [code]
Junle Liu, Yun Zhang, Zixi Guo
ACM TOMM, 2025

Abstract

Feature Coding for Machines (FCM) aims to compress intermediate features effectively for remote intelligent analytics, which is crucial for future intelligent visual applications. In this paper, we propose a Multiscale Feature Importance-based Bit Allocation (MFIBA) for end-to-end FCM. First, we find that the importance of features for machine vision tasks varies with the scales, object size, and image instances. Based on this finding, we propose a Multiscale Feature Importance Prediction (MFIP) module to predict the importance weight for each scale of features. Secondly, we propose a task loss-rate model to establish the relationship between the task accuracy losses of using compressed features and the bitrate of encoding these features. Finally, we develop a MFIBA for end-to-end FCM, which is able to assign coding bits of multiscale features more reasonably based on their importance. Experimental results demonstrate that when combined with a retained Efficient Learned Image Compression (ELIC), the proposed MFIBA achieves an average of 38.202% bitrate savings in object detection compared to the anchor ELIC. Moreover, the proposed MFIBA achieves an average of 17.212% and 36.49% feature bitrate savings for instance segmentation and keypoint detection, respectively. When the proposed MFIBA is applied to the LIC-TCM, it achieves an average of 18.103$\%$, 19.866% and 19.597% bit rate savings on three machine vision tasks, respectively, which validates the proposed MFIBA has good generalizability and adaptability to different machine vision tasks and FCM base codecs.

Framework Overview

Requirements

🧩 This project was trained and tested with:

  • 🐍 Python 3.8

📦 To install required packages, simply run:

pip install -r fasterrcnn_getw/requirements.txt

📊 Dataset

In this work, we employed COCO2017 dataset.

COCO2017 Directory Structure

├── annotations
  ├── instances_train2017.json
  ├── instances_val2017.json
  ├── panoptic_train2017.json
  ├── panoptic_val2017.json
  ├── person_keypoints_train2017.json
  └── person_keypoints_val2017.json
├── train
└── val

For simplicity, we recommend precoding multiscale features through different codecs and storing the results in designated directories. Example path: /home/ta/liujunle/sda2/ELIC/features_for_elic/.

Coding Results Directory Structure

├── features_for_elic
  ├── feature_decode_1
  ├── feature_decode_2
  ├── feature_decode_3
  ...
├── kp_features_for_elic
└── mask_features_for_elic

Test

Object detection

 python /home/ta/liujunle/sda2/fasterrcnn_getw/train.py --data-path path-to-coco --dataset coco --num-classes 90 --model resnet50 --batch-size 16 --pretrained --test-only --lmbda_for_update 1000 

Instance segmentation

 python /home/ta/liujunle/sda2/fasterrcnn_getw/train_maskrcnn.py --data-path path-to-coco --dataset coco --num-classes 90 --model mask_rcnn --batch-size 16 --pretrained --test-only --lmbda_for_update 1000 

Keypoint detection

 python /home/ta/liujunle/sda2/fasterrcnn_getw/train_keypoint.py --data-path path-to-coco --dataset coco_kp --num-classes 90 --model keypoint_rcnn --batch-size 16 --pretrained --test-only --lmbda_for_update 1000 

📖 Citation

If you find our work useful or relevant to your research, please kindly cite our paper:

@article{10.1145/3748654,
author = {Liu, Junle and Zhang, Yun and Guo, Zixi and Huang, Xiaoxia and Jiang, Gangyi},
title = {Multiscale Feature Importance-based Bit Allocation for End-to-End Feature Coding for Machines},
year = {2025},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
issn = {1551-6857},
url = {https://doi.org/10.1145/3748654},
doi = {10.1145/3748654},
note = {Just Accepted},
journal = {ACM Trans. Multimedia Comput. Commun. Appl.},
month = aug,
keywords = {Feature coding for machines, deep learning, image coding, bit allocation, object detection}
}

About

MFIBA: Multiscale Feature Importance-based Bit Allocation for End-to-End Feature Coding for Machines

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Contributors 2

  •  
  •  

Languages